Decentralized GNSS Architecture

Updated 28 December 2025

Decentralized GNSS architecture is a distributed network system that leverages inter-node consensus and local estimation to achieve accurate positioning, navigation, and timing without central fusion.
It employs advanced techniques such as gradient tracking consensus, diffusion adaptation, and Kalman filtering to optimize both timing stability and positional accuracy across spaceborne and terrestrial networks.
Its robust design minimizes communication overhead, scales efficiently with increasing node counts, and enhances resilience against link or node failures through multi-round consensus and error compensation.

A decentralized GNSS (Global Navigation Satellite System) architecture replaces centralized fusion or control elements with distributed, networked estimation and consensus. Individual nodes—satellites, receivers, or atomic clocks—compute local observables and exchange compact messages over time-varying graphs, achieving network-wide positioning, navigation, and timing (PNT) self-organization. Recent progress in decentralized optimization, inter-satellite ranging, consensus mechanisms, and clock synchronization enables full-scale GNSS service autonomously in space and on the ground, with performance and resilience that match or exceed centralized benchmarks.

1. System Models and Network Topologies

Decentralized GNSS architectures instantiate either (i) space-segment networks among LEO satellites, (ii) terrestrial reference-station networks, or (iii) hybrid structures operating at various physical and protocol layers.

Spaceborne LEO GNSS Networks.

Each LEO satellite $l$ maintains a local state $x_l \in \mathbb{R}^{6+2+2F}$ , comprising position increment $\Delta p_l$ , velocity increment $\Delta v_l$ , receiver clock bias/drift $\Delta t_l, \Delta \dot{t}_l$ , carrier-phase and code biases per receiver frequency.
The shared variable $z \in \mathbb{R}^{G(2+2F)}$ collects all GNSS-satellite clock and hardware states across $G$ satellites.
The network is modeled as a time-varying undirected graph $G(t) = (V, E(t), W(t))$ , where $W(t)$ is a doubly stochastic Metropolis weight matrix (Liu et al., 24 Dec 2025).

Global Terrestrial Reference Networks.

The receiver infrastructure is modeled as a time-varying directed graph $G(t) = (\mathcal V, \mathcal E(t))$ with $R$ nodes. Mixing weights $W(t)$ are learned or scheduled to optimize information diffusion, respecting instantaneous communication constraints (Zheng et al., 21 Dec 2025).
UAV swarms and atomic clock ensembles are modeled as fully connected synchronous broadcast networks or as arbitrary graph topologies for ground fiber-connected MACs (Dev et al., 1 Aug 2025, Chen et al., 23 Apr 2025).

2. Distributed Estimation and Consensus Algorithms

Spaceborne LEO GNSS: Gradient Tracking Consensus

The global estimation objective is $\min_{\{x_l\}, z}\sum_{l=1}^L f_l(x_l, z)$ , with per-node weighted least-squares losses $f_l$ .
Distributed optimization proceeds via momentum-accelerated gradient tracking (GT): each node iteratively
1. Performs a heavy-ball momentum step on $z_l^k$ .
2. Applies gradient descent using its local gradient tracker $g_l^k$ .
3. Mixed $\mathcal{K}$ -round consensus among neighbors via repeated averaging with weight matrix $W(t)$ .
4. Solves for $x_l^{k+1}$ in closed form via local weighted least-squares.
5. Updates the gradient-tracking variable using new gradients and neighbor mixing (Liu et al., 24 Dec 2025).
Convergence is linear to the centralized WLS optimum under strong connectivity and Lipschitz assumptions.

Terrestrial Reference and Clock Networks: Diffusion Adaptation

Each node maintains a local estimate $\theta_i^k$ and a gradient-tracking auxiliary $y_i^k$ .
Iterative updates:

$\begin{aligned} \theta_i^{k+1} &= \sum_{j \in \mathcal N_i^{(k \bmod \tau)}} W_{ij}^{(k \bmod \tau)} \theta_j^k - \alpha \nabla f_i(\theta_i^k) \ y_i^{k+1} &= \sum_{j \in \mathcal N_i^{(k \bmod \tau)}} W_{ij}^{(k \bmod \tau)} y_j^k + \nabla f_i(\theta_i^{k+1}) - \nabla f_i(\theta_i^k) \end{aligned}$

Mixing matrices $W^{(\ell)}$ are jointly optimized over time-windows to minimize contraction factor $\epsilon_\tau$ and maximize information flow (Zheng et al., 21 Dec 2025).

Raft-style and Kalman Consensus for PNT Replacement

UAV swarms use Raft consensus with sensor fusion of GNSS/INS/ranging, leader election, log replication, median-based recovery, and majority voting to protect against GNSS denial or adversarial faults (Dev et al., 1 Aug 2025).
Atomic clock networks employ second-order stochastic state-space models, distributed Kalman filtering, and consensus control to synchronize local clocks, with supervisor-driven anchoring or optimal floating control to minimize Allan variance over short and long horizons (Chen et al., 23 Apr 2025).

3. Information Exchange, Bandwidth, and Robustness

Inter-Node Communication

Each satellite or node exchanges $\mathbb{R}^{\text{dim}(z)}$ vectors (momentum+descent variables and gradient trackers) per iteration. For LEO GNSS, $\text{dim}(z)\sim 200$ –300 for GPS L1/L2 (Liu et al., 24 Dec 2025).
In diffusion networks, only per-node state vectors and gradient trackers are exchanged with direct neighbors; aggregation to all-to-all is never required (Zheng et al., 21 Dec 2025).

Efficiency and Scalability

Multi-round consensus with $\mathcal{K}=20$ dramatically reduces disagreement in each iteration, allowing large step sizes and fast convergence.
Compared to centralized fusion, which would require streaming all raw observables, decentralized approaches reduce communication by over an order of magnitude (Liu et al., 24 Dec 2025).
Quantization and error feedback can further compress exchanged vectors with minimal loss (Liu et al., 24 Dec 2025).

Robustness to Link and Node Failures

Missing updates are compensated by last-known values, asynchronous variants preserve convergence, and network partitioning allows continued operation within the largest connected component (Liu et al., 24 Dec 2025).
Packet loss and variable latency are smoothed by multi-round consensus and momentum. In consensus protocols, majority and median rules guarantee resilience to up to $f \leq (n-1)/2$ faulty nodes (Dev et al., 1 Aug 2025).

4. Performance Metrics and Analytical Guarantees

Convergence and Accuracy

For LEO GNSS network GT:
- Standalone orbit RMS error: 2.95 m
- Network (float ambiguities): 0.12 m
- Network (fixed ambiguities): 0.06 m
- Standalone timing RMS error: 7.95 ns
- Network (float): 0.21 ns
- Network (fixed): 0.11 ns (Liu et al., 24 Dec 2025)

Architecture	Orbit RMS Error	Timing RMS Error
Standalone	2.95 m	7.95 ns
Network (float ambiguities)	0.12 m	0.21 ns
Network (fixed ambiguities)	0.06 m	0.11 ns

Global-scale ground reference diffusion matches centralized BLU estimator within $7 \times 10^{-9}$ m for position and $5 \times 10^{-9}$ m for satellite biases (Zheng et al., 21 Dec 2025).
UAV SwarmRaft architecture maintains $1.0 \pm 0.2$ m MAE even with $f = 3$ faulty nodes, compared to $8.2 \pm 1.5$ m for GNSS-only in the same N=10 scenario (Dev et al., 1 Aug 2025).

Convergence Rate

Mean-square deviation to the centralized solution decays linearly (geometric rate) in both moment-GT and diffusion architectures, with contraction constant set by network spectral properties and optimization parameters (Liu et al., 24 Dec 2025, Zheng et al., 21 Dec 2025).

5. Relativistic Autonomous Architectures

The Autonomous Basis of Coordinates (ABC) approach constructs a fully decentralized GNSS reference frame using only inter-satellite signal exchanges and emission coordinates (Kostić et al., 2014). The key components are:

All satellites continuously exchange proper-time beacons and solve the null-geodesic (“light-cone”) equations, yielding emission coordinates $(\tau_1, \tau_2, \tau_3, \tau_4)$ at every event.
The network jointly refines orbital parameters using action minimization over residual light travel times, absorbing clock offsets and detecting perturbations.
The result is a millimetre-level inertial reference frame, completely decoupled from terrestrial reference and limited primarily by satellite clock noise and on-board gravitational modeling.

6. Distributed Atomic Timing and Time-Scale Construction

Atomic clock networks utilize hierarchical, peer-to-peer protocols to construct global time-scales with optimal short- and long-term Allan variance (Chen et al., 23 Apr 2025).

In normal mode, periodic GNSS anchoring aligns the ensemble mean with the standard time; in emergency (GNSS-outage) mode, floating optimal control minimizes divergence using only local measurements.
Distributed Kalman filtering on graph edge-states, consensus control, and explicit weighting vector selection jointly ensure that both short- and long-term frequency stability targets are achieved, with no single point of failure.

7. Implications, Scalability, and Integration Outlook

Decentralized GNSS architectures eliminate single-point failure risks and bottlenecks associated with centralized processing hubs. They scale naturally with node count; each satellite, station, or clock ensemble requires only $O(1)$ neighbor interactions and modest local memory and compute resources (e.g., <10 MB RAM per LEO satellite for full orbit/clock estimation (Liu et al., 24 Dec 2025)). Architectural robustness, resilience to communication outages, and rapid convergence are achieved at communication costs far below those of centralized streaming fusion.

A plausible implication is that, as LEO and mixed-orbit mega-constellations proliferate, fully networked, self-organizing PNT will become the operational default, with user-facing GNSS services delivered through autonomous, in-orbit and ground-based peer-to-peer fusion. Similarly, decentralized atomic timing and SwarmRaft-style consensus mechanisms will underlie emerging PNT services in challenging, adversarial, or degraded-signal environments.

Key references: